Computer pioneer Alan Turing’s remarks in 1950 on the question, “Can machines think?” were misquoted, misinterpreted and morphed into the so-called “Turing Test”. The modern version says if you can’t tell the difference between communicating with a machine and a human, the machine is intelligent. What Turing actually said was that by the year 2000 people would be using words like “thinking” and “intelligent” to describe computers, because interacting with them would be so similar to interacting with people. Computer scientists do not sit down and say alrighty, let’s put this new software to the Turing Test - by Grabthar’s Hammer, it passed! We’ve achieved Artificial Intelligence!
I think the Chinese room argument published in 1980 gives a pretty convincing reason why the Turing test doesn’t demonstrate intelligence.
The thought experiment starts by placing a computer that can perfectly converse in Chinese in one room, and a human that only knows English in another, with a door separating them. Chinese characters are written and placed on a piece of paper underneath the door, and the computer can reply fluently, slipping the reply underneath the door. The human is then given English instructions which replicate the instructions and function of the computer program to converse in Chinese. The human follows the instructions and the two rooms can perfectly communicate in Chinese, but the human still does not actually understand the characters, merely following instructions to converse. Searle states that both the computer and human are doing identical tasks, following instructions without truly understanding or “thinking”.
Searle asserts that there is no essential difference between the roles of the computer and the human in the experiment. Each simply follows a program, step-by-step, producing behavior that makes them appear to understand. However, the human would not be able to understand the conversation. Therefore, he argues, it follows that the computer would not be able to understand the conversation either.
I am sceptical of this thought experiment as it seems to imply that what goes on within the human brain is not computable. For reference: every single physical effect that we have thus far discovered can be computed/simulated on a Turing machine.
The argument itself is also riddled with vagueness and handwaving: it gives no definition of understanding but presumes it as something that has a definite location, and also it may well be possible that taking the time to run the program inevitably causes understanding of Chinese after even the first word returned. Remember: executing these instructions could take billions of years for the presumably immortal human in the room, and we expect the human to be so thorough that they execute each of the trillions of instructions without error.
Indeed, the Turing test is insufficient to test for intelligence, but the statement that the Chinese room argument tries to support is much, much stronger than that. It essentially argues that computers can’t be intelligent at all.
That just shows a fundamental misunderstanding of levels. Neither the computer nor the human understands Chinese. Both the programs do, however.
The programs don’t really understand Chinese either. They are just filled with an understanding that is provided to them up-front. I mean as in they do not derive that understanding from something they perceive where there was no understanding before, they don’t draw conclusions, don’t understand words from context,… the way an intelligent being would learn a language.
Nothing in the thought experiment says that the program doesn’t behave that way. If the program really seems like it understands language to an outside observer, you would assume it did learn language that way.
Others have provided better answers than mine, pointing out that the Chinese room argument only makes sense if your premise is that a “program” is qualitatively different from what goes on in a human brain/mind.
Programs clearly understand words from context. Try making it do translation tasks, it can properly translate “tear” to either 泪水 (tears from crying) or 撕破 (to rend) based on context
The problem with the experiment is that there exists a set of instructions for which the ability to complete them necessitates understanding due to conditional dependence on the state in each iteration.
In which case, only agents that can actually understand the state in the Chinese would be able to successfully continue.
So it’s a great experiment for the solipsism of understanding as it relates to following pure functional operations, but not functions that have state changing side effects where future results depend on understanding the current state.
There’s a pretty significant body of evidence by now that transformers can in fact ‘understand’ in this sense, from interpretability research around neural network features in SAE work, linear representations of world models starting with the Othello-GPT work, and the Skill-Mix work where GPT-4 and later models are beyond reasonable statistical chance at the level of complexity for being able to combine different skills without understanding them.
If the models were just Markov chains (where prior state doesn’t impact current operation), the Chinese room is very applicable. But pretty much by definition transformer self-attention violates the Markov property.
TL;DR: It’s a very obsolete thought experiment whose continued misapplication flies in the face of empirical evidence at least since around early 2023.
It was invalid when he originally proposed it because it assumes a unique mystical ability for the atoms that make up our brains. For Searle the atoms in our brain have a quality that cannot be duplicated by other atoms simply because they aren’t in what he recognizes as a human being.
It’s why he claims the machine translation system system is incapable of understanding because the claim assumes it is possible.
It’s self contradictory. He won’t consider it possible because it hasn’t been shown to be possible.
The Chinese room experiment only demonstrates how the Turing test isn’t valid. It’s got nothing to do with LLMs.
I would be curious about that significant body of research though, if you’ve got a link to some papers.
Searle argued from his personal truth that a mystic soul is responsible for sapience.
His argument against a computer system having consciousness is this:
" In order for this reply to be remotely plausible, one must take it for granted that consciousness can be the product of an information processing “system”, and does not require anything resembling the actual biology of the brain."
-Searle
Isn’t the brain just an information processing system?
Brilliant thought experiment. I never heard of it before. It does seem to describe what’s happening - if only there were a way to turn it into a meme so modern audiences could understand it.
I can’t remember who said this, but somebody said the version of the Turing Test as we all remember it is ridiculous: It’s basically saying that the test of intelligence is “Can a chatbot fool one idiot?”
That’s essentially the media-generated Turing Test, but in truth no such test was ever defined by Alan Turing. For me the modern takeaway is don’t extrapolate anything about reality from memes.
More “can fool the average idiot.”
‘Passing’ isn’t fooling a single participant, but the majority of them beyond statistical chance.
This is a good discussion about the original tests Turing discussed, as well as a very fun interview.
I always saw it more as pragmatism relating to humanity and being possibly extended to machine intelligence by association. When you talk with another person you have no real way of knowing that they are separate conscious entities, intelligent and self aware in the way you perceive yourself to be. But if they talk and act in a way that is suggestive of that then the best and simplest working practice is to assume it. This same practicality should extend to include artificial intelligence as applicable.
Yes I think that’s generally what Alan Turing meant - he was careful not to define what “intelligence” means, and was discussing practical perception of machine behavior.
The Turing Test codified the very real fact that computer AI systems up till a few years ago couldn’t hold a conversation (outside of special conversational tricks like Eliza and Cleverbot). Deep neural networks and the attention mechanism changed the situation; it’s not a completely solved problem, but the improvement is undeniably dramatic. It’s now possible to treat chatbots as a rudimentary research assistant, for example.
It’s just something we have to take in stride, like computers becoming capable of playing Chess or Go. There is no need to get hung up on the word “intelligence”.
Not sure how you define getting “hung up” but there are tons of poorly informed people who believe/fear that AI is about to take over/conquer/destroy/whatever the world because they think LLMs are as smart as humans - or just a few tweaks away. It’s less about the word “intelligence” than about jumping from there to collateral issues, like thinking LLMs are “persons” that deserve rights, that using them without their consent is slavery, and other nonsense. Manipulative people take advantage of this kind of ignorance. Knowledge is good, modern superstition is bad.
They are going to destroy the world, not because they are superintelligent but because LLMs will be linked to lethal weapons and critical machines since it’s easier to learn than a human and since they are very not reliant (prompt injection, purposely lying, etc.), this will lead to death
It’s just something we have to take in stride, like computers becoming capable of playing Chess or Go. There is no need to get hung up on the word “intelligence”.
Nicely said.
Y’all might enjoy reading Blindsight. Really digs into questions of sapience, intelligence, etc. Is it evolutionary cost worth it? I’ve read it 15+ times. Because I’m a psycho.
“You think we’re nothing but a Chinese Room,” Rorschach sneered. “Your mistake, Theseus.”
And suddenly Rorschach snapped into view—no refractory composites, no profiles or simulations in false color. There it was at last, naked even to Human eyes.
Imagine a crown of thorns, twisted, dark and unreflective, grown too thickly tangled to ever rest on any human head. Put it in orbit around a failed star whose own reflected half-light does little more than throw its satellites into silhouette. Occasional bloody highlights glinted like dim embers from its twists and crannies; they only emphasized the darkness everywhere else.
Imagine an artefact that embodies the very notion of torture, something so wrenched and disfigured that even across uncounted lightyears and unimaginable differences in biology and outlook, you can’t help but feel that somehow, the structure itself is in pain.
Now make it the size of a city.
The Turing Test as it is popularly conceptualized is really more of a test of human intelligence (or stupidity, more likely) rather than that of the machine.
If you put a big enough idiot in front of the screen, Dr. Sbaitso could conceivably “pass.” Well, maybe if you muted it, anyway.
Eliza, a chatbot psychiatry emulator written in the 1960s, convinced many people it was a real person.
Various versions of Eliza are online - including this quaint, retro looking one
